Long-horizon tasks that require sustained reasoning and multiple tool interactions remain challenging for LLM agents: small errors compound across steps, and even state-of-the-art models often hallucinate or lose coherence.We identify context management as the central bottleneck-extended histories cause agents to overlook critical evidence or become distracted by irrelevant information, thus failing to replan or reflect from previous mistakes.To address this, we propose COMPASS (Context-Organized Multi-Agent Planning and Strategy System), a lightweight hierarchical framework that separates tactical execution, strategic oversight, and context organization into three specialized components: (1) a Main Agent that performs reasoning and tool use, (2) a Meta-Thinker that monitors progress and issues strategic interventions, and (3) a Context Manager that maintains concise, relevant progress briefs for different reasoning stages.Across three challenging benchmarks-GAIA, BrowseComp, and Humanity's Last Exam-COMPASS improves accuracy by up to 20% relative to both single-and multi-agent baselines.We further introduce a test-time scaling extension that elevates performance to match established DeepResearch agents, and a posttraining pipeline that delegates context management to smaller models for enhanced efficiency.

The rise of artificial intelligence (AI) in medical diagnostics has highlighted an essential need for transparent and interpretable systems, particularly in the field of mental health. The opaque decision-making in “black-box” AI models creates a challenge in clinical settings: they risk losing trust or causing uncritical reliance. This paradox jeopardizes mental healthcare, where skepticism or unwarranted confidence in AI can harm patient care. Effective explainability mechanisms are therefore essential not only to earn trust but to support responsible use by allowing clinicians to critically assess and verify AI-driven outputs. Thus, we propose a novel explainable AI (XAI) pipeline for automated depression diagnosis, designed to integrate both system-level and human-level interpretability. This dual approach is vital in clinical settings, as it combines rigorous statistical validation with clear, actionable insights that enhance practitioner confidence in AI-generated diagnoses. The pipeline leverages deep learning models for classification, augmented by traditional system-level XAI techniques and a Retrieval-Augmented Generation (RAG)-enhanced Large Language Model (LLM). With the integration of LLMs, the system translates abstract system-level explanations into understandable, natural language narratives. This provides a crucial cross-verification step that fosters calibrated trust: it mitigates the dual risks of under-trust, by providing a clear rationale, and over-trust, by flagging diagnostic inconsistencies within the AI models. This capability is critical for ensuring both user trust and satisfaction, as it empowers practitioners to critically assess and validate AI-driven insights. Through comprehensive human evaluations conducted by medical professionals, this approach demonstrates high alignment with clinical diagnostic indicators, underscoring the value of combining system-level and human-level explanations to make complex AI processes transparent and clinically meaningful. Validated across three diverse datasets-KangNing, EDAIC-WOZ, and CALLM-each presenting unique structural challenges from structured clinical inquiries to narrative-driven dialogues, our method bridges the gap between technical AI outputs and practitioner understanding, marking a significant advancement toward a trust-based, widely adoptable AI diagnostic tool in mental health care. By introducing a comprehensive framework for combining statistical rigor with narrative clarity, this work represents a critical step toward closing the gap between black-box AI systems and real-world clinical adoption. While our study is limited to depression diagnosis, the proposed framework illustrates a pathway toward explainable clinical AI systems. Future work may explore its adaptability to other medical contexts.

Causal discovery (CD) and Large Language Models (LLMs) have emerged as transformative fields in artificial intelligence that have evolved largely independently. While CD specializes in uncovering cause-effect relationships from data, and LLMs excel at natural language processing and generation, their integration presents unique opportunities for advancing causal understanding. This survey examines how LLMs are transforming CD across three key dimensions: direct causal extraction from text, integration of domain knowledge into statistical methods, and refinement of causal structures. We systematically analyze approaches that leverage LLMs for CD tasks, highlighting their innovative use of metadata and natural language for causal inference. Our analysis reveals both LLMs' potential to enhance traditional CD methods and their current limitations as imperfect expert systems. We identify key research gaps, outline evaluation frameworks and benchmarks for LLM-based causal discovery, and advocate future research efforts for leveraging LLMs in causality research. As the first comprehensive examination of the synergy between LLMs and CD, this work lays the groundwork for future advances in the field.

Guangya Wan, Yuqi Wu, Jie Chen, Sheng Li. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.